Skip to content

CLAUDE.md

Latest commit

 

History

History
67 lines (50 loc) · 2.43 KB

CLAUDE.md

File metadata and controls

67 lines (50 loc) · 2.43 KB

CLAUDE.md - Project Guidelines for Claude Code

Project Overview

cell2location is a Bayesian model for spatial transcriptomics that estimates cell type abundances in spatial data by integrating single-cell RNA-seq reference signatures. It is built on top of scvi-tools and Pyro.

Repository: https://github.com/BayraktarLab/cell2location

Tech Stack

  • Language: Python 3.10+
  • Core dependencies: pyro-ppl, scvi-tools, torch, numpy, pandas, scanpy
  • ML framework: PyTorch + Pyro (probabilistic programming)
  • Data format: AnnData (h5ad files)

Code Style & Formatting

  • Black for code formatting (line length: 120)
  • isort for import sorting (profile: black, trailing comma)
  • flake8 for linting (ignored: E203, E266, E501, W503, W605, N812; max line length: 119)
  • Pre-commit hooks enforce all of the above

Run formatting before committing:

black --line-length 120 .
isort .
flake8

Project Structure

cell2location/
├── models/           # Main model classes (Cell2location, RegressionModel)
│   ├── base/         # Base modules and mixins (Pyro integration)
│   ├── reference/    # Reference signature estimation model
│   └── simplified/   # Simplified model variants
├── nn/               # Neural network layers (FC layers, context layers)
├── dataloaders/      # Custom data loaders for spatial data
├── distributions/    # Custom probability distributions
├── cell_comm/        # Cell communication analysis
├── cluster_averages/ # Cluster average computation
├── plt/              # Plotting utilities
└── utils/            # General utilities

Testing

Tests are in tests/. Run with:

pytest

CI runs on Python 3.10 (Ubuntu). Tests include model training smoke tests.

Key Architecture Patterns

  • Models follow the scvi-tools pattern: a user-facing Model class wraps a PyTorch Module
  • Pyro is used for variational inference (SVI with ELBO)
  • QuantileMixin and PltExportMixin add posterior quantile computation and plotting
  • Data registration uses scvi-tools AnnDataManager with field types (LayerField, CategoricalObsField, etc.)

Important Notes

  • Always ensure adata.var_names matches cell_state_df.index when using Cell2location
  • The N_cells_per_location parameter significantly affects model behavior
  • Detection mean correction handles sensitivity differences between spatial and single-cell data